[Showoff Saturday] I built a open-source playground for testing LLM APIs and MCP tools in the browser

I’ve been building an open-source developer playground for experimenting with LLM APIs and remote MCP servers from one web interface. The motivation was simple: besides my desire to use cheaper models with GitHub MCP and Tavily for coding (so I can make an agent change some files even from mobile), I wanted to see the actual requests, streamed responses, tool arguments, token estimates, and latency without switching between several scripts and dashboards. I also wanted every tool execution to remain explicitly controlled by the user. The project currently supports: • Custom OpenAI-compatible endpoints • Remote MCP servers over Streamable HTTP or SSE • A built-in preset for GitHub’s hosted MCP server • Manual approval or denial before any MCP tool call is executed • A manual tool runner for invoking individual MCP tools with JSON arguments • Raw request and response inspection • Streaming responses and reasoning display • TypeScript and curl snippet export • Locally stored chats, settings, prompt templates, and optional API keys using IndexedDB • Automatic expiry of remembered API keys after 30 days GitHub: https://github.com/ownyourwebsite/ai-playground Live demo: https://ai-playground.ownyourwebsite.app/ The code is MIT licensed.